Anthropic has published a new set of measurements designed to show how quickly artificial intelligence is becoming involved in the process of developing more artificial intelligence.
The company says the public has limited visibility into what happens inside frontier AI laboratories. Its new framework attempts to make that process easier to track by measuring three areas: how much AI performs AI research and development, how AI agents are monitored, and how computing resources are allocated.
The measurements are particularly notable because they provide an internal snapshot of how Anthropic is using Claude in its own model-development work. As of August 2026, Anthropic says Claude was leading 26% of its measured AI R&D tasks, while more than 90% of the work was performed at a level where AI at least collaborated with humans.
What Is Anthropic Measuring?
Anthropic's framework focuses on the production process behind advanced AI systems rather than simply measuring what the models can do.
The company has proposed three main measurements:
- AI-led AI R&D — how much of the work involved in developing AI is itself performed by AI.
- AI agent oversight — how thoroughly actions taken by AI agents are monitored and how quickly potentially problematic activity is reviewed.
- Compute allocation — how much computing capacity is being directed toward AI research and safety-related work.
Anthropic says these measurements could eventually give researchers, governments and the public a clearer view of how rapidly frontier AI development is progressing.
The approach is also connected to a broader question: how much of the work required to create increasingly capable AI systems can eventually be automated?
Claude Is Already Leading a Significant Share of AI R&D
One of the most striking figures in Anthropic's report concerns its AI R&D Automation Index.
The company uses an automation scale ranging from AL0 to AL5. At the lower end, AI has little or no involvement. At AL3, AI collaborates with humans. At AL4, AI leads a task and can complete most of it from a high-level instruction while a human supervises. AL5 represents full autonomy without a human in the loop.
Anthropic's August 2026 snapshot found:
- Claude was not fully autonomous for any measured category of AI R&D.
- Claude was classified as leading 26% of Anthropic's AI R&D work.
- More than 90% of the measured work was at or above the AI-collaboration level.
That represents a significant change from earlier in the year. Anthropic's published chart shows the AL4 share was below 1% in February 2026 before reaching 26% in August.
The numbers do not mean that Claude is independently running Anthropic's entire research organization. Instead, they indicate the estimated share of defined R&D tasks where AI can now perform most of the work with humans providing high-level supervision.
Why AI-Led R&D Matters
AI systems helping engineers write code or analyze research papers is not entirely new. The more important shift is when AI becomes capable of contributing directly to the development of future AI systems.
This creates a potential feedback loop.
A model can help researchers design experiments, diagnose problems, improve training processes, evaluate other models and develop software. Those improvements can contribute to more capable future models, which can then perform more sophisticated research tasks.